‘From resistance to expressive drive’: how a domain-specific LLM enhances L2 writing outcomes through guided feedback

Generic large language models (LLMs), including ChatGPT, have recently been applied in second language writing instruction due to their capability to provide immediate feedback on students’ writing. However, these models still exhibit limitations, including ‘hallucinations’ and overcorrection. In response, we developed Dr. Write, a domain-specific AI-powered writing feedback system trained on the world’s largest corpus of German writing by Chinese learners. This study outlines the pedagogical rationale behind the design of Dr. Write and empirically evaluates its effectiveness through a quasi-experimental design. A total of 124 L2 German learners were assigned to three feedback conditions: a domain-specific LLM (Dr. Write), a generic LLM (Qwen), and teacher feedback. The results showed that both GAI feedback conditions outperformed teacher feedback in learners’ cognitive and behavioural engagement with feedback, but not in affective engagement. Dr. Write was also associated with significant gains in writing self-efficacy, particularly in self-regulatory efficacy and performance self-efficacy, as well as in clause-level writing accuracy. Qualitative findings further suggest that its guided, proficiency-aligned feedback fostered active revision, self-monitoring, and reflection. The study highlights the pedagogical potential of domain-specific LLMs for L2 writing and offers a transferable framework for future research on AI-supported feedback in educational settings.